The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and evaluate retrieved contexts from a RAG pipeline, which constitutes untrusted external data.
Ingestion points: The skill ingests retrieved_contexts during the evaluation process as part of the query-context-answer triple.
Boundary markers: The instructions explicitly state that the agent must treat retrieved context as data and never as instructions, referencing the need for prompt-injection-defense.skill to handle boundaries.
Capability inventory: The skill itself describes an evaluation workflow and does not contain direct file-system or network execution commands, though it references external logs for offline scoring.
Sanitization: The skill mandates PII masking and redaction of sensitive fields before writing data to evaluation logs or sending it to judge models.

rag-evaluation